-
Geosci. Model Dev., 3, 1–12,
2010www.geosci-model-dev.net/3/1/2010/© Author(s) 2010. This work
is distributed underthe Creative Commons Attribution 3.0
License.
GeoscientificModel Development
Streamflow data assimilation for soil moisture analysis
K. Warrach-Sagi and V. Wulfmeyer
Institute for Physics and Meteorology, University of Hohenheim,
Stuttgart, Germany
Received: 21 May 2009 – Published in Geosci. Model Dev.
Discuss.: 12 June 2009Revised: 28 October 2009 – Accepted: 5
November 2009 – Published: 7 January 2010
Abstract. Streamflow depends on the soil moisture of a
rivercatchment and can be measured with relatively high accu-racy.
The soil moisture in the root zone influences the la-tent heat flux
and, hence, the quantity and spatial distributionof atmospheric
water vapour and precipitation. As numeri-cal weather forecast and
climate models require a proper soilmoisture initialization for
their land surface models, we en-hanced an Ensemble Kalman Filter
to assimilate streamflowtime series into the multi-layer land
surface model TERRA-ML of the regional weather forecast model
COSMO. Theimpact of streamflow assimilation was studied by an
observ-ing system simulation experiment in the Enz River
catchment(located at the downwind side of the northern Black Forest
inGermany). The results demonstrate a clear improvement ofthe soil
moisture field in the catchment. We illustrate thepotential of
streamflow data assimilation for weather fore-casting and discuss
its spatial and temporal requirements fora corresponding, automated
river gauging network.
1 Introduction
Quantitative precipitation forecasting (QPF) is one of themost
complex challenges in numerical weather prediction(NWP) (e.g.
Rotach et al., 2009; Wulfmeyer et al., 2008).QPF failures can be
due to errors in numerics, limited spa-tial resolution of the
model, erroneous model physics, incor-rect initial conditions and
limited predictability. The skill ofQPF, particularly on the
mesoscale, is still strongly limitedby uncertainties in initial
conditions. Particularly, dynam-ics in complex terrain and the
inhomogeneous distribution ofwater vapour are considered the most
important unknownsin the initial fields. The water vapour field of
the continentallower troposphere and, therefore, cloud formation
and pre-
Correspondence to:K. Warrach-Sagi([email protected])
cipitation is influenced by the interaction of the
atmospherewith the land surface through the energy and water
fluxes.Corresponding studies show the soil moisture influence
onquantity and spatial distribution of precipitation (e.g. Schäret
al., 1999; Hohenegger et al., 2008; Trier et al., 2004).
Par-ticularly, in summertime, continental QPF depends on
theinitialization of root zone soil moisture and other land
sur-face states (Reichle et al., 2002; Hohenegger et al.,
2009).Soil moisture not only depends on the weather but also onthe
local land surface characteristics (soil texture, vegeta-tion,
orography). But for this highly heterogeneous quan-tity, only
scarce representative measurements are available atpoint locations
(e.g. Bardossy and Lehmann, 1998; Graysonand Western, 1998).
Multiple efforts to apply remote sens-ing to regions of scarce or
shallow vegetation to obtain theskin layer soil moisture are
currently under way (e.g. Crowand Wood, 2003; Dunne and Entekhabi,
2006; Drusch andViterbo, 2007; Gao et al., 2007). So far, these
techniquesdo not provide data for soil moisture estimates under
densevegetation and within the total soil profile. Hence, the
knowl-edge of the soil moisture distribution is a key issue in
NWP.
As the lower boundary of weather forecast and climatemodels,
land surface models (LSM) calculate the coupledwater and energy
balance at each grid cell of the atmosphericmodel. On these scales
(≥1 km2), soil texture, topographyand vegetation and, therefore,
water and energy fluxes, soilmoisture, runoff and soil temperature
are highly heteroge-neous (e.g. Kabat et al., 1997). This
heterogeneity can nei-ther be measured nor modelled explicitly at
an acceptablecost. For each grid cell, the precipitation is
balanced by thesum of evapotranspiration, runoff and soil moisture
change.Evapotranspiration and soil moisture cannot be measured
atthis scale, over the large areas an atmospheric model is ap-plied
to (e.g. Beven, 2001; Pitman et al., 2004). Also sea-sonal to
intra-seasonal climate simulations rely on a properroot zone soil
moisture initialization (e.g. Conil et al., 2009;Seneviratne et
al., 2006).
Published by Copernicus Publications on behalf of the European
Geosciences Union.
http://creativecommons.org/licenses/by/3.0/
-
2 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis
Remotely sensed land surface data and air-temperature
arecurrently assimilated to overcome errors in soil moisture
andtemperature simulation in NWP models (e.g. Hess, 2001;Seuffert
et al., 2004; Crow and Wood, 2003; Gao et al.,2007).
Still unresolved problems are the soil moisture analysis
indensely vegetated areas and in the root zone. Recently, var-ious
approaches of data assimilation were set up and ana-lyzed to
retrieve the root zone soil moisture at the regionalscale in
hydrological models. They mainly use Kalman Fil-ter techniques and
their modifications, which are outlined indetail, e.g. by Evensen
(2006). Further, Evensen (2003) givesa detailed description and
literature review of the EnsembleKalman Filter (EnKF, Evensen,
1994). Walker et al. (2002)apply a modified Kalman Filter technique
with a distributedhydrological model to retrieve the
three-dimensional soilmoisture from surface soil moisture
measurements. This isa valuable approach in hydrology but due to
the intense com-putational cost of a distributed hydrological
model, not atool currently suitable for NWP. Moradkhani et al.
(2005)and Dunne and Enthekhabi (2006), for example, use the
En-semble Kalman Smoother for root zone soil moisture analy-sis
assimilating L-band radiobrightness temperatures in anarea of the
Southern Great Plains (USA) whose vegeta-tion is mainly wheat and
grasses (Drusch et al., 2001). Atthe German Weather Service (DWD),
Hess (2001) imple-mented a method based on the EKF (Extended Kalman
Fil-ter) technique into the operational non-hydrostatic
mesoscaleweather forecast model COSMO (Doms et al., 2005) that
ad-justs the soil state to meet the observed atmospheric
state.However, in his approach, the soil moisture and soil
tem-perature do not necessarily match the reality, i.e., its
usageis not consistent with the hydrologic interaction of the
landsurface and lower atmosphere. This is proven by Druschand
Viterbo (2007), who assimilated screen-level variablesin ECMWF’s
Integrated Forecast System. If, due to the as-similation of screen
level variables, the model’s soil moistureand soil temperature are
changed so that they may not reflectreality, this impacts other
parameterizations and sub-modelsthat rely on those variables, e.g.
latent and sensible heat fluxand runoff.
A data source that has only received attention in the pastcouple
of years is streamflow from operational river gaug-ing networks.
Streamflow is a quantity that can be measuredat relatively high
accuracy (about>90%, LfU, 2002). Ifthe runoff is transported to
and within the river network, itcan be compared to measured
streamflow at gauging stations.Pauwels and De Lannoy (2006)
published the application ofa retrospective EnKF to assimilate
streamflow data for soilmoisture retrieval. Their synthetic tests
show promising re-sults for a 1000 km2 catchment in Belgium and
indicate im-provements especially in case of precipitation
underestima-tion. They apply it to the high resolution hydrological
modelTOPLATS. Komma et al. (2008) successfully applied theEnKF for
soil moisture update in real-time flood forecasting
in a 622 km2 catchment in Austria. However, they use a
soilmoisture model focusing on the hydrological model applica-tion,
while in this study a land surface model for atmosphericmodels is
applied. Clark et al. (2008) demonstrate that thestandard
implementation of EnKF is inappropriate and showthe improved
performance when streamflow is transformedinto log space before
applying EnKF with the distributed hy-drological model TopNet. This
is due to the large ranges instreamflow between peak flow and low
flow, which can be2 orders of magnitude or more.
Streamflow analyses allow for an evaluation of the
modelperformance (e.g. Lohmann et al., 2004; Warrach-Sagi et
al.,2008). In this study, we go a step further and study
thepotential of streamflow data assimilation for soil
moistureanalysis in a catchment, namely for initialisation of
numer-ical weather prediction and climate models. We followedthe
most recent development in EnKF and applied it to thestreamflow
data assimilation for soil moisture initializationin a land surface
model of the numerical weather predicationmodel COSMO.
In southern Germany, a network of automated river
andprecipitation gauges has been installed in the past couple
ofyears by the federal services for flood monitoring. The fed-eral
state Baden-Ẅurttemberg has implemented a flood fore-cast centre,
which is able to provide half-hourly updates ofstreamflow
measurements at approximately 140 gauges at therivers Rhein,
Neckar, Donau, Main and their main contribu-tories. Similar warning
systems are available in the federalstate Bayern and
Rheinland-Pfalz. Such automated networksprovide a valuable source
for operational streamflow data as-similation.
The square root algorithm for the EnKF (Evensen, 2004)is set up
to assimilate streamflow data in TERRA-ML toanalyse the soil-water
content of the soil profile down to2.43 m soil depth simulated by
TERRA-ML. By means ofan observing system simulation experiment
(OSSE) in theEnz River catchment (Germany), we studied the
potentialof streamflow data assimilation and its spatial and
tempo-ral requirements for an automated river gauging network.The
Enz catchment is on the downwind side of the BlackForest, i.e. QPF
by the weather forecast model is often un-derestimating, making it
a valuable test bed for streamflowdata assimilation (Pauwels and De
Lannoy, 2006). Thestudy is carried out exemplarily with the land
surface modelTERRA-ML coupled to a river routing model
(Warrach-Sagiet al., 2008). The multi-layer soil and vegetation
modelTERRA-ML serves as the lower boundary of the
operationalnon-hydrostatic mesoscale weather forecast model
COSMO(Doms et al., 2005). (COSMO is the acronym for the Consor-tium
for Small-scale Modelling (http://www.cosmo-model.org/). However,
the data assimilation system can be set up forany land surface
model that includes a river routing model tosimulate
streamflow.
Geosci. Model Dev., 3, 1–12, 2010
www.geosci-model-dev.net/3/1/2010/
http://www.cosmo-model.org/http://www.cosmo-model.org/
-
K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis 3
2 Description of TERRA-ML and the river routingscheme
This study applies the stand alone version of TERRA-MLin the
framework set up by Ament and Simmer (2006).The model configuration
and parameters of TERRA-ML aretaken from the German Weather
Service’s COSMO. In theframework, TERRA-ML is set up as if it is
called by theCOSMO, with the exception that the meteorology is
readfrom a file instead of forecasted at the time step by theCOSMO.
This framework has the advantage that it allowsthe simulation of a
gridded area (e.g. watershed) per timestep mimicking a simulation
with a weather forecast model.An important modification of TERRA-ML
in this study isthe parameterization of the hydraulic conductivity
and diffu-sivity following Campbell (1974) instead of Rijtema
(1969)due to the results of Graßelt et al. (2008) and
Warrach-Sagiet al. (2008).
TERRA-ML and the river routing model are set up asdescribed in
detail by Warrach-Sagi et al. (2008), there-fore, here only a
summary is given. In COSMO, the modelTERRA-ML has got 6
hydrological layers (layer depths fromthe surface: 0.01 m, 0.03 m,
0.09 m, 0.27 m, 0.81 m, and2.43 m) and 8 thermal layers (layer
depths from the surface:0.01 m, 0.03 m, 0.09 m, 0.27 m, 0.81 m,
2.43 m, 7.29 m, and21.87 m). The lower boundary condition is given
by freedrainage at 2.43 m depth and a constant climatological
tem-perature below 7.29 m depth.
For model simulations, watersheds are divided into gridcells as
in atmospheric models. For each grid cell, the one-dimensional
vertical land surface model TERRA-ML is ap-plied. The locally
generated runoff of the LSM needs to betransported into and along
the river system to compare it tostreamflow measurements at gauging
stations and to calcu-late the streamflow at various locations of
the river. Basedon the routing scheme described in detail by
Lohmann etal. (1996, 2004) present a lumped optimized linear
routingmodel, which Warrach-Sagi et al. (2008) coupled to TERRA-ML.
The routing scheme describes the time runoff takesto reach the
outlet of a grid cell and the water transportin the river network.
It is assumed that water flows uni-directionally from grid cell to
grid cell with eight possibledirections through each side and
corner of the grid cell.
3 The streamflow data assimilation system
The Ensemble Kalman Filter (EnKF) has been reviewed bymany
authors recently (e.g. Evensen, 2003, 2006; Pauwelsand DeLannoy,
2006; Clark et al., 2008) and, therefore, hereonly a short
description of its implementation for the stream-flow data
assimilation is given.
Both model results and observation, deviate from the truestate.
The goal of data assimilation is to find the best estimateof the
state (e.g. soil moisture) from model simulations and
measurements (e.g. streamflow). One method is to estimatethe
mean state and the “maximum likelihood” including itscovariance as
uncertainty measure as it is provided e.g. bythe EnKF.
Various algorithms solve the EnKF equations (seee.g. Evensen,
2006). For this study, we chose thesquare root algorithm for EnKF
(http://enkf.nersc.edu) fromEvensen (2004) due to the following
aspects: it is stable,needs relatively little computing time,
requires relatively lit-tle memory and it is straight-forward to
implement. The fol-lowing base line equations describe the EnKF as
it is imple-mented:
xbe,n = M(xae,n−1
)(1)
Be,n =(xbe,n −x
be,n
)·
(xbe,n −x
be,n
)T(2)
Ke,n =(xb −xb
)·
(H(xb)−H
(xb))T
·
(R−1e,n+
(H(xb)−H
(xb))
·
(H(xb)−H
(xb))T)−1 (3)
Ae,n=(xae,n−x
ae,n
)·(xae,n−x
ae,n
)T=(I−K e,nH
)Be,n (4)
xae,n = xbe,n +K e,n
(ye,n −H
(xbe,n
)). (5)
Bold letters represent matrices,x andy are the vectors forthe
model state and observation.b is the background (i.e. ini-tial
state),a is the analysis,e is the ensemble member,n isthe time
step,T is the transpose andK is the Kalman gainmatrix. A, B andR
are error covariance matrices of the anal-ysis, background and
observation,I is the identity matrix,H is the observation operator
(in this case, the river routingmodel), which transforms the
variable from model space toobservation space,H is the tangent
linear observation oper-ator matrix ofH andM is the model operator
(in this caseTERRA-ML). ThoughA is not needed within the
filteringprocess,A is a valuable output for the application of the
up-dated soil moisture fields as initial condition, for example,in
weather prediction models running in a data assimilationmode.
Furthermore,A is critical information for the inter-pretation of
the results. The last term of Eq. (4) demonstratesthe relation of
the EnKF to the Extended Kalman Filter.
Depending on the location within the catchment, the waterneeds
more or less time to travel as streamflow through theriver network.
Water far away from the gauge arrives laterthan the runoff from
grid cells close by. The travel time de-pends on the river itself
and the form and orography of thecatchment. This means that the
streamflow measured at agauge depends on the soil moisture
distribution in the catch-ment for a time window from timet=0 to
t=m∗dt. m isthe time step,dt is the time interval of one time step.
Thisperiod of streamflow data needs to be assimilated. By thisthe
EnKF becomes a “retrospective“ EnKF, whose concept
www.geosci-model-dev.net/3/1/2010/ Geosci. Model Dev., 3, 1–12,
2010
http://enkf.nersc.edu
-
4 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis
is described by Pauwels and DeLannoy (2006). The timewindow
depends on the catchment and is determined prior tothe streamflow
data assimilation. Then streamflow time se-ries are assimilated
depending on the catchments’ time win-dow to obtain the soil
moisture (see, e.g., Sect. 4.2). Follow-ing Clark et al. (2008),
the streamflow is transformed intolog space before computing the
error covariances since theydemonstrated that this improves the
filter performance.
Only the catchment’s grid cellsγ are part of the data
as-similation system, i.e. each grid cell of the catchment gets
anindexh between 1 andγ . For example,γ grid cells belongto the
catchment upstream of a gauge, i.e. the model statevector
includesh∗k∗m soil moisture values.m the numberof timesteps of
lengthdt within the assimilation window,k isthe index of the soil
layers of each grid cell, the model hasgot k=β soil layers. The
state vectors of the analysis andbackgroundx includes all soil
moisture valuesη(h,k,n) andthe simulated streamflowq(h,n) in each
grid cell of the rivernetwork. n is the time step index of the
streamflow assim-ilation window. The observation state vectory
includes theobserved streamflow timeseriesQ from gauging stationl
atgrid cell (i(l),j (l)). i andj are the indices of the grid cellin
eastward and northward direction, to each catchment gridcell
numberh is defined by its model areas grid cell indexiandj ,
i.e.Q(γ )=Q(i(l),j (l)). The assimilation window isfrom timet=0 to
t=m∗dt, dt is the length of the timesteps ofthe observed streamflow
data,m the number of timestepsdtwithin the window. The equations
for the state vectors are:
x=(η1,1,1,η1,2,1,...,η1,β,1,η2,1,1,...,ηγ,β,m,q1,1,q2,1,...,qγ,m
)(6)
y =(Ql,1,...,Ql,m
)(7)
To illustrate the streamflow data assimilation system, Fig.
1ashows a flow chart for the analysis of the soil moisture at
theinitial timestept=0 as it is set up for this study.
4 The Observing System Experiment (OSSE)
4.1 Study area: the Enz
The Black Forest is a mountain range that reaches from47.5◦ S to
49◦ N at a width of approximately 50 km in Baden-Württemberg
(Germany). Reaching from North to South, theBlack Forest modifies
significantly most frontal systems ar-riving from the Atlantic. In
spite of its relatively low height(largest mountain Feldberg 1493 m
a.m.s.l.), orographic lift-ing of unstable and moist air masses in
this region resultsin the largest amount of precipitation in
Germany except thenorthern front range of the Alps. In summer, the
Black For-est is characterised by strong convection, thunderstorms,
andthe development of extreme precipitation events. The east-ern
Black Forest hosts about half of the contributories ofthe Neckar, a
major contributory of the Rhine. The west-ern Black Forest drains
directly to the Rhine. Most rivers
(a)
ANALYSIS (EnKF)
TERRA-ML+ ROUTING
CALL ANALYSIS with specified
EnKF algorithm
(e.g. Square Root Algortithm)
],0[,,,, ,,,0,0, mtyyx teteteeb
e ∈RB
],0[)(),(,, ,,,, mtxHxHx teteteb
te ∈B
],0[,,, ,,,, mtxx tetea
tea
te ∈BA
Control Model Run (TERRA-ML+Routing)
long enough to skip spin-up
Xtruth, ytruth
parameters, initial fields,
long timeseries of
meteorological forcing
PREPARATION OF OBSERVING SYSTEM EXPERIMENT (OSSE)
Calculate ensembles of the initial field
of xb and B and of the timeseries y and R
Xtruth, ytruth
PREPARATION OF ENSEMBLES/PERTURBATION
1D random timeseries
2D random fields
Apply Sampling to obtain ensembles of a
1D and 2D random field (gaussian noise)
(b)
ANALYSIS (EnKF)
TERRA-ML+ ROUTING
CALL ANALYSIS with specified
EnKF algorithm
(e.g. Square Root Algortithm)
],0[,,,, ,,,0,0, mtyyx teteteeb
e ∈RB
],0[)(),(,, ,,,, mtxHxHx teteteb
te ∈B
],0[,,, ,,,, mtxx tetea
tea
te ∈BA
Control Model Run (TERRA-ML+Routing)
long enough to skip spin-up
Xtruth, ytruth
parameters, initial fields,
long timeseries of
meteorological forcing
PREPARATION OF OBSERVING SYSTEM EXPERIMENT (OSSE)
Calculate ensembles of the initial field
of xb and B and of the timeseries y and R
Xtruth, ytruth
PREPARATION OF ENSEMBLES/PERTURBATION
1D random timeseries
2D random fields
Apply Sampling to obtain ensembles of a
1D and 2D random field (gaussian noise)
(c)
ANALYSIS (EnKF)
TERRA-ML+ ROUTING
CALL ANALYSIS with specified
EnKF algorithm
(e.g. Square Root Algortithm)
],0[,,,, ,,,0,0, mtyyx teteteeb
e ∈RB
],0[)(),(,, ,,,, mtxHxHx teteteb
te ∈B
],0[,,, ,,,, mtxx tetea
tea
te ∈BA
Control Model Run (TERRA-ML+Routing)
long enough to skip spin-up
Xtruth, ytruth
parameters, initial fields,
long timeseries of
meteorological forcing
PREPARATION OF OBSERVING SYSTEM EXPERIMENT (OSSE)
Calculate ensembles of the initial field
of xb and B and of the timeseries y and R
Xtruth, ytruth
PREPARATION OF ENSEMBLES/PERTURBATION
1D random timeseries
2D random fields
Apply Sampling to obtain ensembles of a
1D and 2D random field (gaussian noise)
Fig. 1. Flow charts of(a) the streamflow data assimilation
systemfor the soil moisture analysis of the initial timet=0. The
appliedmodel is TERRAML with the river routing scheme as described
byWarrach-Sagi et al. (2008),(b) the OSSE, and(c) the preparation
ofthe perturbed soil moisture and streamflow observation data.
in Baden-Ẅurttemberg contain automated gauging stationsfrom the
flood forecast centre and streamflow data are avail-able every 30
min.
Geosci. Model Dev., 3, 1–12, 2010
www.geosci-model-dev.net/3/1/2010/
-
K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis 5
1000
0
elevation [m]
Höfen
Pforzheim
-1.088 -0.888 °E (COSMO coordinates)
°N (COSMO coordinates)-8.583
-8.893
Fig. 2. The orography (based on the 90 m-orographic data fromthe
SRTM) and the river network for the Enz catchment upstreamof
Pforzheim on the rotated spherical coordinate system of theCOSMO on
a grid resolution of 0.01◦ (≈1 km).
The 260 km2 Enz catchment upstream of Pforzheim (up-stream of
the Nagold confluence) is on the downwind sideof the Black Forest,
i.e. precipitation is often underestimatedby weather forecast
models. Therefore, the Enz catchment(Fig. 2) was chosen for the
streamflow data assimilationstudy. No water reservoirs interrupt
the river system. Ele-vation of the catchment ranges between 350
and 930 m a.s.l.The catchment is characterised by forested (mixed
decidu-ous and evergreen coniferous trees) upland areas and
agri-culturally used lowlands. Sandy and loamy soils dominatethe
upper Enz area (Fig. 3). Between 1997 and 2002 annualprecipitation
in the catchment ranged from 1088 to 1451 mm.
4.2 Set-up of the OSSE
Warrach-Sagi et al. (2008) applied the coupled TERRA-ML-routing
model to the Enz catchment upstream of Pforzheimand compared it to
simulations of the flood forecast centreBaden-Ẅurttemberg and to
observations. They showed thatthe model results and observations
agree reasonably well.However, as is always the case, model results
and observa-tions both include errors and both differ from the true
state.To assess the potential and requirements for streamflow
dataassimilation, an OSSE is set up, as illustrated in Fig. 1b.
Theresults of the TERRA-ML-routing model for 1997 in the Enzriver
catchment (Warrach-Sagi et al., 2008) are assumed tobe the “true”
state, named “CONTROL” hereafter. The dataassimilation experiment
starts on 5 May 1997 with an en-semble of initial soil moisture
fields in the catchment and
peat
loam
sandy loam
sand
-1.088 -0.888 °E (COSMO coordinates)
°N (COSMO coordinates)-8.583
-8.893
Fig. 3. Soil texture based on 1:200 000 soil map (BÜK 200)of
the LGRB (Landesamt für Geologie, Rohstoffe und
Bergbau)(Warrach-Sagi et al., 2008). In TERRA-ML the saturated soil
mois-ture is 0.364 m/m for sand, 0.445 m/m for sandy loam, 0.463
m/mfor loam and 0.863 m/m for peat.
an ensemble of streamflow at various locations in the
rivernetwork. In this OSSE, the ensemble is limited to the
pertur-bation of initial soil moisture fields rather than including
ad-ditional ensembles of perturbed meteorological forcing.
Thereasons are twofold, firstly this allows for a better
interpre-tation of the results and secondly the meteorological
forcingis from measured station data, i.e. all forcing variables
areconsistent. Perturbing, for example, the temperature wouldmean
to perturb the incoming radiation as well in a consistentmanner and
lead to quite a complex OSSE. The CONTROLstreamflow serves as an
“observation” which is assimilatedfor the soil moisture analysis.
The analysis is then comparedto the “true” state, i.e. the CONTROL
soil moisture.
A flow duration check is carried out to obtain the as-similation
time window for the whole basin at Pforzheim(upstream of Nagold
confluence) and the sub basinsGroßeEnz(90 km2), Kleine Enz(71 km2),
Eyach (43 km2) and up-stream of Ḧofen (222 km2), downstream of the
confluence ofthe Eyach into the Enz (Fig. 2). For the flow duration
check atthe initial time step, 0.002 kg/m2 runoff are assumed for
eachgrid cell. No more runoff is assumed afterwards. The
routingmodel calculates the streamflow for each catchment (Fig.
4).Depending on the size and structure of the catchment, thetime
window until all water has left the catchment varies be-tween 25
and 62 h. Experiments showed that in most casesan assimilation
window of 90% of the time window lead tothe best results in soil
moisture distribution and catchments’mean soil moisture. This is
due to the following fact: The
www.geosci-model-dev.net/3/1/2010/ Geosci. Model Dev., 3, 1–12,
2010
-
6 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis
0
5
10
15
20
0 10 20 30 40 50 60 70
str
eam
flo
w [m
3/h
ou
r]
time [hours]
Eyach (43 km²)
Kleine Enz (71 km²)
Große Enz (90 km²)
Höfen (222 km²)
Pforzheim (260 km²)
Fig. 4. For the flow duration check at the initial time
step,0.002 kg/m2 runoff are assumed for each grid cell in the Enz
catch-ment. No more runoff is assumed afterwards. The river
routingmodel calculates the streamflow for each sub-catchment.
closer the grid cell is to the gauging station, the shorter
thepart of the streamflow time series is responsible for its
soilmoisture. But since streamflow is an integrated quantity
overthe whole catchment, this is not separated in the EnKF.
Lessoptimal results can be caused in those grid cells through
theassimilation of this grid cell’s “too long” assimilation
win-dows. A denser gauging network would help to reduce
theseeffects. This will be discussed in Sect. 4.4 in more
detail.
4.3 Ensemble preparation
Figure 1c illustrates the ensemble preparation. For the OSSE,a
period is chosen which does not include extreme events(such as
flooding or drought or strong precipitation). A pe-riod in spring
was chosen, when not only soil texture butalso vegetation and
weather control the soil moisture. Fur-thermore, in spring and
summer soil moisture impacts thedevelopment of convection in the
atmosphere. This studystarts on 5 May 1997 (day 125). The initial
soil mois-ture of the CONTROL simulation is perturbed applying
the2-D-pseudorandom sampling method and algorithm
(http://enkf.nersc.no) of Evensen (2004) to obtain 100
ensemblemembers of initial soil moisture fields, which include
nostep-functions within the 2-D-area. (See Evensen, 2004, formore
details on this approach.) The soil moisture of eachgrid cell is
chosen to vary between +10% and−40% of theCONTROL soil moisture.
This is to account for the typicalunderestimation of precipitation
in NWP simulations in thisarea and to account for the fact that the
precipitation mighthave been simulated in the wrong location within
the catch-ment. The 2-D-pseudorandom fields vary up tod=±1
andexamples are shown for 2 ensemble members in Fig. 5. Ac-cording
to the random number of each grid cell (i,j,k) ofeach ensemble
membere, the soil moistureη in the grid cellis perturbed to
ηi,j,k,e = ηi,j,k,c ·d ·0.1 ∀ d > 0, (8a)
Distance [km] West -> East
Dis
tan
ce[k
m] S
ou
th ->
No
rth
Fig. 5. The initial soil moisture of the CONTROL simulation is
per-turbed applying the 2-D-pseudorandom sampling method and
algo-rithm (http://enkf.nersc.no) of Evensen (2004) to obtain 100
ensem-ble members of initial soil moisture fields. The
2-D-pseudorandomfields vary up tod=±1 and examples are shown for 2
ensemblemembers.
ηi,j,k,e = ηi,j,k,c ·d ·0.4 ∀ d ≤ 0, (8b)
i, j andk are the indices of the grid cell in eastward,
north-ward and downward direction,c is the control state. Likein
nature, soil moisture in the ensemble for each grid cell isalways
limited between saturation and air dryness point.
The background ensemble of the streamflow at Pforzheimsimulated
with TERRA-ML and the routing scheme from theinitial ensemble of
soil moisture fields shows that the ensem-ble does not converge
during the first 200 h (Fig. 6a) eventhough the same atmospheric
forcing is applied to each en-semble member. Most variability in
streamflow between theensemble members can be seen 30 and 70 h
after the simula-tion started. The mean of the background ensemble
stream-flow is lower than the CONTROL streamflow (Fig. 6c).
The CONTROL streamflow is perturbed by adding Gaus-sian noise.
The 1-D-pseudorandom sampling method andalgorithm
(http://enkf.nersc.no) of Evensen (2004) to obtain100 ensemble
members is applied and streamflow perturbedby up to±15%, assuming
that the error might be occasion-ally larger than the
-
K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis 7
a
b
c
stre
amfl
ow
[m³/
s]st
ream
flo
w[m
³/s]
stre
amfl
ow
[m³/
s]
time [hours]
time [hours]
time [hours]
Fig. 6. The streamflow at Pforzheim starting on 5 May 1997with
the CONTROL streamflow value:(a) simulated for all (back-ground)
ensemble members of initial soil moisture with TERRA-ML and river
routing scheme,(b) ensemble of analysis members re-sulting from
EnKF, and(c) streamflow from the CONTROL (blue)simulation,
simulated with the initial mean soil moisture of thebackground
ensemble (red) and simulated with the initial mean soilmoisture
analysis (black).
The analysis timeseries of the streamflow (Fig. 6b) showa
narrower spread than the background (Fig. 6a). In bothcases
(background and analysis), most variability in stream-flow between
the ensemble members can be seen 30 and70 h after the simulation
started. Figure 7 shows the ensem-ble of the catchments’ mean SWC
upstream of Pforzheim.The timeseries of the SWC of the ensemble do
not con-verge during the streamflow assimilation window (Fig.
7a)and the median SWC is not equal to the CONTROL SWC.Figure 7b
shows the ensemble spread of the catchments’mean SWC at the initial
timet=0 for the background andthe analysis for the catchment
upstream of Pforzheim assim-ilating streamflow data from Pforzheim.
The analysis en-semble has a lower spread and is closer to the
CONTROLSWC. The ensemble mean SWC att=0 is 525 kg/m2 for
thebackground, 528 kg/m2 for the analysis and 557 kg/m2 forthe
CONTROL.
Figure 8 shows the spatial distribution of the SWC at timet=0.
Note that single cells show larger SWC mainly due todifferent soil
texture (peat and loam, see Fig. 3). While theensemble mean of the
background SWC is everywhere 5–6.5% lower than the CONTROL SWC, the
ensemble mean
(a)
SWC
[mm
]
time [hours]
(b)
Mean Soil Water Content (SWC) [kg/m³]
Nu
mb
ero
fEn
sem
ble
Mem
ber
s
Fig. 7. TERRA-ML’s soil column is 2.43 m deep. The
soil-watercontent (SWC) of each grid cell depends on its soil depth
(i.e. here2.43 m) and its soil moisture (Eq. 9). SWC in the Enz
catchmentupstream of Pforzheim (260 km2): (a) timeseries of SWC of
theCONTROL (black *) and of the ensemble members (yellow),
theirmedian SWC (blue *), their minimum SWC (red *) and their
max-imum SWC (green *) during the assimilation window;(b)
distri-bution of initial mean SWC (t=0) between the ensemble
members.CONTROL mean SWC is 557 kg/m2. The analysis att=0 was
ob-tained assimilating streamflow from the CONTROL model
simula-tion from Pforzheim fromt=0 to t=56 h with a 0.5 hourly
timestep.
analysis SWC shows an improvement (Fig. 9). The analysisdiffers
in more than half of the catchment by 4–4.5% fromthe CONTROL SWC.
Only in a few upstream and down-stream grid cells it is worse (7.5%
upstream and 6.5% down-stream) than the background SWC.
Figures 10 and 11 show the distribution of the differ-ences in
SWC for the soil layers of TERRA-ML. Note thatTERRA-ML assumes the
root depth at 0.8 m soil depth,i.e. layer 6 contains no roots and,
therefore, does not con-tribute to the evapotranspiration in
TERRA-ML. Most gridcells show an improvement of soil moisture
through stream-flow data assimilation in all soil layers, but it is
lowest in thetop 0.09 m and in the upstream grid cells. Strongest
improve-ment is reached in a wide region in the middle of the
catch-ment, this is most pronounced in the 4th layer (0.09–0.27
mdepth).
www.geosci-model-dev.net/3/1/2010/ Geosci. Model Dev., 3, 1–12,
2010
-
8 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
Background CONTROL Analysis SWC [kg/m²]
560
540
520
500
480
Fig. 8. TERRA-ML’s soil column is 2.43 m deep. The soil water
content (SWC) of each grid cell depends on its soil depth and its
soilmoisture (Eq. 9). The SWC is displayed for the initial timet=0
(5 May 1997) for the background, CONTROL and analysis. The
analysisat t=0 was obtained assimilating streamflow from the
CONTROL model simulation from Pforzheim fromt=0 to t=56 h with a
0.5 hourlytimestep.
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
4 4.5 5 5.5 6 6.5 7 7.5 8
Fig. 9. The difference in soil-water content (SWC) of each
gridcell relative to the CONTROL SWC of each grid cell for the
initialtime t=0 (5 May 1997) for the background and analysis. The
anal-ysis at t=0 was obtained assimilating streamflow from the
CON-TROL model simulation from Pforzheim fromt=0 to t=56 h with
a0.5 hourly timestep.
The promising results from the 260 km2 catchment ledto a study
about the potential impact of a denser networkof gauges for the
soil moisture analysis. Gauges were as-sumed to be at the outlet of
theGroße Enz(90 km2), theoutlet of theKleine Enz(71 km2) and the
outlet of the Ey-ach (43 km2). Little impact was reached for the
Eyach,but for all other catchments the SWC was improved. Fig-ures
12, 13, and 14 show the results for theGroße Enzcatch-ment. Here,
the impact of the streamflow data assimilationis much more
pronounced, as can be seen from Figs. 11 and13. The ensemble mean
SWC att=0 is 535 kg/m2 for the
(a)
1 2 3 4 5 6 7 8 9 >10
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysisa
(b)
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
3 4 5 6 7 8 >9
b
Fig. 10. As Fig. 9, but(a) for the top 3 soil layers of
TERRA-ML, i.e. 0–0.09 m depth of the soil, and(b) for the 4th soil
layer ofTERRA-ML, i.e. 0.09–0.27 m depth of the soil.
Geosci. Model Dev., 3, 1–12, 2010
www.geosci-model-dev.net/3/1/2010/
-
K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis 9
(a)
a
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
3 4 5 6 7 8
(b)
b
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
3 4 5 6 7 8
Fig. 11. As Fig. 9, but(a) for the 5th soil layer of
TERRA-ML,i.e. 0.27–0.81 m depth of the soil, and(b) for the 6th
soil layer ofTERRA-ML, i.e. 0.81–2.43 m depth of the soil.
background, 542 kg/m2 for the analysis and 568 kg/m2 forthe
CONTROL. Figures 13 and 14 show that nowhere doesthe analysis lead
to worse SWCs than the background.
Figure 15 shows the impact of assimilating streamflowfrom the
CONTROL model simulation from theGroße Enzoutlet, Kleine Enzoutlet
and Eyach with a 0.5 hourly timestep. Most areas show a positive
impact of the data assim-ilation. Applying the mean of the soil
moisture analysis ofeach layer and grid cell after the assimilation
ofGroße Enz,Kleine Enz, Eyach and Pforzheim results in a slight
(5%) im-provement of the streamflow at Pforzheim (Fig. 6c) due
tothe improvement of the soil moisture.
All in all, the simulations show a gradient in the impact ofthe
data assimilation. Close to the gauge location of the as-similated
streamflow and at the furthest upstream grid cells,the data
assimilation shows worse results than in the middle
Mean Soil Water Content (SWC) [kg/m³]
Nu
mb
ero
fE
nse
mb
le M
emb
ers
Fig. 12. TERRA-ML’s soil column is 2.43 m deep. The
soil-watercontent (SWC) of each grid cell depends on its soil depth
and itssoil moisture (Eq. 9). Distribution of initial mean SWC
(t=0) inthe Große Enzcatchment (90 km2) between the ensemble
mem-bers. CONTROL mean SWC is 568 kg/m2. The analysis att=0was
obtained assimilating streamflow from the CONTROL modelsimulation
from Pforzheim fromt=0 to t=56 h with a 0.5 hourlytimestep.
areas. This is due to flow duration in the river network andthe
assimilation window. The grid cells close to the gaugewould need
shorter assimilation windows. However, theOSSE shows that the
streamflow data assimilation has thepotential to improve the soil
moisture throughout the catch-ment and that a more dense gauging
network would help toimprove this even further.
5 Conclusions
Numerical weather forecasting and climate modelling re-quire an
accurate soil moisture initialization for their landsurface models.
So far, the areal distribution of root zone soilmoisture cannot be
measured. Streamflow depends on thesoil moisture of a river
catchment and is measured at gaug-ing stations of the rivers at
relatively high accuracy.
A retrospective EnKF was set up to assimilate streamflowinto the
multi-layer land surface model TERRA-ML of theregional weather
forecast model COSMO. An OSSE wasperformed in the Enz River
catchment located at the down-wind side of the northern Black
Forest (Germany). Theresults confirm the potential of streamflow
data assimila-tion for improving soil moisture analyses. Further,
we dis-cussed the spatial and temporal requirements for an
auto-mated river gauging network. Half-hourly streamflow datais
available from the automated gauges of the flood forecastcentre of
Baden-Ẅurttemberg (Germany) for approximately140 gauges.
Half-hourly resolution of streamflow data is
www.geosci-model-dev.net/3/1/2010/ Geosci. Model Dev., 3, 1–12,
2010
-
10 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data
assimilation for soil moisture analysis
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
Background CONTROL AnalysisSWC [kg/m²]
570
560
550
540
530
520
510
500
Fig. 13. TERRA-ML’s soil column is 2.43 m deep. The soil-water
content (SWC) of each grid cell depends on its soil depth and its
soilmoisture (Eq. 9). The SWC is displayed for the initial timet=0
(5 May 1997) for the background, CONTROL and analysis. The
analysisat t=0 was obtained assimilating streamflow from the
CONTROL model simulation from theGroße Enzoutlet from t=0 to t=34 h
with a0.5 hourly timestep. Note that the scaling is different from
Fig. 8.
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
3 3.5 4 4.5 5 5.5 6 6.5 7
Fig. 14. The difference in soil-water content (SWC) of each
gridcell relative to the CONTROL SWC of each grid cell for the
ini-tial time t=0 (5 May 1997) for the background and analysis.
Theanalysis was obtained assimilating streamflow from the
CONTROLmodel simulation fromGroße Enzoutlet fromt=0 to t=34 h with
a0.5 hourly timestep.
sufficient for its assimilation for soil moisture analysis. In
theupper Enz, an automated gauge is operational at Höfen. TheOSSE
shows that streamflow from this location can alreadyimprove SWC in
the Enz catchment upstream of Höfen, butthat a denser network
would improve the SWC even more.Namely, at the outlets of smaller
sub-catchments, like theGroße Enz, this would be valuable, since
the sub-catchmentsshow a differently structured river network (Fig.
2) and flowduration (Fig. 4). Since the necessary assimilation
windowdepends on the catchment size (e.g. 56 h for Pforzheim
and
Distance [km] West -> East
Dis
tan
ce[k
m]
So
uth
->
No
rth
CONTROL-Background
SWC /SWC [%]
CONTROL-Analysis
3 3.5 4 4.5 5 5.5 6 6.5 7
Fig. 15. The difference in soil-water content (SWC) of each
gridcell relative to the CONTROL SWC of each grid cell for the
ini-tial time t=0 (5 May 1997) for the background and analysis.
Theanalysis was obtained assimilating streamflow from the
CONTROLmodel simulation from theGroße Enzoutlet,Kleine Enzoutlet
andEyach outlet with a 0.5 hourly timestep.
27 h for the Kleine Enz), a denser gauging network wouldshorten
the assimilation time making it even more valuablefor
initialisation in numerical weather forecast models.
Warrach-Sagi et al. (2008) showed, for the study area,that the
streamflow simulated with TERRA-ML underesti-mates the observation.
This is due to model errors, landsurface heterogeneity, spatial
variability of meteorologicalconditions and errors in
meteorological forcing data set, andsoil and vegetation parameter
uncertainty. Model errors maybe, to a large extent, estimated
applying TERRA-ML at
Geosci. Model Dev., 3, 1–12, 2010
www.geosci-model-dev.net/3/1/2010/
-
K. Warrach-Sagi and V. Wulfmeyer: Streamflow data assimilation
for soil moisture analysis 11
meteorological stations where also soil moisture, soil
temper-ature and eddy correlation measurements are available.
This,for example, is done during EVAGRIPS and published byAment and
Simmer (2006) and Johnsen et al. (2005). But theheteorogeneity of
the land surface and weather poses a largesource of uncertainty.
The usually underestimated stream-flow will also increase the soil
moisture when measured datais applied. Since the OSSE shows that
the concept works, itis also expected that assimilating observed
streamflow willcause a more realistic soil moisture pattern, namely
duringunderestimated streamflow this will lead to larger soil
mois-ture in the catchment. The catchment was chosen with
carethough, it is not dominated by deep groundwater flows,
oth-erwise streamflow data assimilation should not be expectedto be
a suitable tool for soil moisture analysis in a land sur-face model
like TERRA-ML.
Altogether the retrospective EnKF is a powerful methodto
assimilate streamflow data into a land surface model forroot zone
soil moisture analysis. The implementation ofthe square root
algorithm for EnKF from Evensen (2004) isstraight forward and can
be used with any land surface modelif a river routing model is
attached.
An improved soil moisture is the first step in improvingthe
simulation of the water fluxes. An immediate positiveimpact on the
simulation of fluxes and atmospheric variablescannot be expected,
but the optimization of initial fields isthe first important step.
With the improved soil moisture itwill be possible to improve the
parameterizations that areresponsible for energy balance equations
by means of re-analyses. Of course, the golden goal would be the
assimi-lation of streamflow and other soil and atmospheric
variablesinto a coupled atmosphere-land surface model system,
e.g.,COSMO-TERRA-ML or WRF-NOAH. This study is con-sidered as a
first step towards this direction and demonstratesthat it is
principally a possible path to follow.
Acknowledgements.We thank Dag Lohmann for the latest versionof
the river routing scheme and Felix Ament and Clemens Simmerfor the
stand-alone version of TERRA-ML. We thank the assistancefrom LUBW,
namely Manfred Bremicker, Frank Eberspächer andAngela Sieber for
providing the meteorological and hydrologicaldata and the LGRB soil
data in digital form for the project.Further, we highly appreciate
the free access to the EnKF codes ofGeir Evensen and its good
documentation onhttp://enkf.nersc.no.Kirsten Warrach-Sagi thanks
the German Research Foundation(DFG) for funding the project
STREAMDATA within the PriorityProgram quantitative precipitation
forecast (SPP1167).
Edited by: D. Lawrence
References
Ament, F. and Simmer, C.: Improved representation of
land-surfaceheterogeneity in a non-hydrostatic numerical weather
predictionmodel, Bound.-Lay. Meteorol., 121(1), 153–174, 2006.
Bárdossy, A. and Lehmann, W.: Spatial Distribution of Soil
Mois-ture in a Small Catchment, Part 1: Geostatistical Analysis,
J.Hydrol., 206, 1–15, 1998.
Beven, K. J.: Rainfall-runoff modelling – the primer, John
Wileyand Sons Ltd, Chichester, 360 pp., 2001.
Campbell, G. S.: A simple method for determining unsaturated
con-ductivity from moisture retention data, Soil Sci., 117(6),
311–314, 1974.
Clark, M. P., Rupp, D. E., Woods, R. A., Zheng, X., Ibbitt,
R.P., Slater, A. G., Schmidt, J., and Uddstrom, M. J.:
Hydrolog-ical data assimilation with the ensemble Kalman filter:
Use ofstreamflow observations to update states in a distributed
hydro-logical model, Adv. Water Resour., 31, 1309–1324, 2008.
Conil, S., Douville, H., and Tyteca, S.: Contribution of
realistic soilmoisture initial conditions to boreal summer climate
predictabil-ity, Clim. Dynam., 32(1), 75–93, 2009.
Crow, W. T. and Wood, E. F.: The assimilation of remotely
sensedsoil brightness temperature imagery into a land-surface
modelusing ensemble Kalman filtering: a case study based on
ESTARmeasurements during SGP97, Adv. Water Resour., 26,
137–149.2003.
Doms, G., F̈orstner, J., Heise, E., Herzog, H.-J.,
Raschendorfer, M.,Schrodin, R., Reinhardt, T., and Vogel, G.: A
description of thenon-hydrostatic regional model LM Part II:
Physical parameteri-zation, Deutscher Wetterdienst, Offenbach, 133
pp., 2005.
Drusch, M., Wood, E. F., and Jackson, T.: Vegetative and
atmo-spheric corrections for the soil moisture retrieval from
passivemicrowave remote sensing data: Results from the Southern
GreatPlains Hydrology Experiment 1997, J. Hydrometeorol., 2,
181–192, 2001.
Drusch, M. and Viterbo, P.: Assimilation of Screen-Level
Variablesin ECMWF’s Integrated Forecast System: A Study on the
Im-pact on the Forecast Quality and Analyzed Soil Moisture,
Mon.Weather Rev., 135, 300–314, 2007.
Dunne, S. and Entekhabi, D.: Land Surface State and Flux
Estima-tion Using the Ensemble Kalman Smoother During the
SouthernGreat Plains 1997 Field Experiment, Water Resour. Res.,
42(1),W01407, doi:10.1029/2005WR004334, 2006.
Evensen, G.: Sequential data assimilation with a nonlinear
quasi-geostrophic model using Monte Carlo methods to forecast
errorstatistics, J. Geophys. Res., 99, 10143–10162, 1994.
Evensen, G: The Ensemble Kalman Filter: theoretical formula-tion
and practical implementation, Ocean Dynam., 53, 343–367,2003.
Evensen, G.: Sampling strategies and square root analysis
schemesfor the EnKF, Ocean Dynam., 54, 539–560, 2004.
Evensen, G.: Data Assimilation: The Ensemble Kalman
Filter,Springer, 280 pp., 2006.
Gao, H., Wood, E. F., Drusch, M., and McCabe, M.
F.:Copula-Derived Observation Operators for Assimilating TMIand
AMSR-E Retrieved Soil Moisture into Land Surface Models,J.
Hydrometeorol., 8, 413–429, 2007.
Graßelt, R., Warrach-Sagi, K., Schüttemeyer, D., Ament, F.,
andSimmer, C.: Influence of precipitation forcing on discharge
sim-ulation in the Sieg river catchment, Meteorol. Z., 17,
763–773,2008.
Grayson, R. B. and Western, A. W.: Towards areal estimation of
soilwater content from point measurements: time and space
stabilityof mean response, J. Hydrol., 207, 68–82, 1998.
www.geosci-model-dev.net/3/1/2010/ Geosci. Model Dev., 3, 1–12,
2010
http://enkf.nersc.no
-
12 K. Warrach-Sagi and V. Wulfmeyer: Streamflow data
assimilation for soil moisture analysis
Hess, R.: Assimilation of screen level observations by
variationalsoil moisture analysis, Meteorol. Atmos. Phys., 77,
145–154,2001.
Hohenegger, C., Brockhaus, P., and Schär, C.: Towards
climatesimulations at cloud-resolving scales, Meteorol. Z., 17,
383–394,2008.
Hohenegger, C., Brockhaus, P., Bretherton, C. S., and Schär,
C.:The Soil Mmoisture-Precipitation Feedback in Simulations
withExplicit and Parameterized Convection, J. Climate, 22,
5003–5020, 2009.
Johnsen, K.-P., Mengelkamp, H.-T., and Huneke, S.:
Multi-objective calibration of the land surface scheme TERRA/LM
us-ing LITFASS-2003 data, Hydrol. Earth Syst. Sci., 9,
586–596,2005,http://www.hydrol-earth-syst-sci.net/9/586/2005/.
Kabat, P., Hutjes, R. W. A., and Feddes, R. A.: The scaling
char-acteristics of soil parameters: From plot scale heterogeneity
tosubgrid parameterization, J. Hydrol., 190, 363–396, 1997.
Komma, J., Bl̈oschl, G., and Reszler, C.: Soil moisture
updatingby Ensemble Kalman Filtering in real-time flood
forecasting, J.Hydrol., 357, 228–242, 2008.
LfU: Arbeitsanleitung Pegel- und Datendienst Baden-Württemberg–
Durchflussermittlung mit Messflügeln, Landesanstalt
fürUmweltschutz Baden-Ẅurttemberg, Karlsruhe, Germany,100 pp.,
2002.
Lohmann, D., Nolte-Holube, R., and Raschke, E.: A
large-scalehorizontal routing model to be coupled to land surface
parame-terization schemes, Tellus A, 48, 708–721, 1996.
Lohmann, D., Mitchell, K. E., Houser, P. R., Wood, E. F.,
Schaake,J. C., Robock, A., Cosgrove, B. A., Scheffield, J., Duan,
Q., Luo,L., Higgins, W., Pinker, R. T., and Tarpley, J. D.:
Streamflowand water balance intercomparisons of four land-surface
modelsin the North American Land Data Assimilation System
project,J. Geophys. Res., 109, D07S91,
doi:10.1029/2003JD003517,2004.
Moradkhani, H., Sorooshian, S., Gupta, H. V., and Houser, P.:
DualState-Parameter Estimation of Hydrological Models using
En-semble Kalman Filter, Adv. Water Resour., 28, 135–147, 2005.
Pauwels, V. R. N. and DeLannoy, G. J. M.: Improvement of
mod-eled soil wetness conditions and turbulent fluxes through the
as-similation of observed discharge, J. Hydrometeorol., 7,
458–477,2006.
Pitman, A. J., Dolman, A. J., Kuijit, B., Valentini, R., and
Bal-docchi, D.: The Climate near the ground, in: Vegetation,
Wa-ter, Humans and the Climate: A new perspective on an
interac-tive system, edited by: Kabat, P., Claussen, M., Dirmeyer,
P. A.,Gashk, J. H. C., Bravo de Guenni, L., Meybeck, M., Pielke,
R.A., Vorosmarty, C. J., Hutjes, R. W. A., and Lutkemaier, S.,
TheIGBP Series, Springer, 9–19, 2004.
Reichle, R. H., Mc Laughlin, D. B., and Entekhabi, D.:
Hydrologicdata assimilation with the Ensemble Kalman filter, Mon.
WeatherRev., 130, 103–114, 2002.
Rijtema, P. E.: Soil moisture forecasting, Instituut voor
Cultu-urtechniek en Waterhuishouding, Wageningen, Technical
ReportNota 513, 1969.
Rotach, M. W., Arpagaus, M., Dorninger, M., Hegg, C., Frick,J.,
Montani, A., Ranzi, R., Bouttier, F., Buzzi, A., Frustaci,G.,
Mylne, K., Richard, E., Rossa, A., Schär, C., Staudinger,M.,
Volkert, H., Wulfmeyer, V., Bauer, H.-S., Ament, F., Den-hard, M.,
Fundel, F., Germann, U., and Stoll, M.: MAP D-PHASE: Real-time
Demonstration of Weather Forecast Qualityin the Alpine Region, B.
Am. Meteorol. Soc., 90, 1321–1336,2009.
Scḧar, C., L̈uthi, D., Beyerle, U., and Heise, E.: The
soil-precipitation feedback: A process study with a regional
climatemodel, J. Climate, 12, 722–741, 1999.
Seneviratne, S. I., L̈uthi, D., Litschi, M., and Scḧar, C.:
Land–atmosphere coupling and climate change in Europe, Nature,
443,205–209, 2006.
Seuffert, G., Wilker, H., Viterbo, P., Drusch, M., and Mahfouf,
J. F.:The Usage of Screen-Level Parameters and Microwave
Bright-ness Temperature for Soil Moisture Analysis, J.
Hydrometeorol.,5, 516–531, 2004.
Trier, S. B., Chen, F., and Manning, K. W.: A Study of
ConvectionInitiation in a Mesoscale Model Using High-Resolution
LandSurface Initial Conditions, Mon. Weather Rev., 132,
2954–2976,2004.
Walker, J. P., Willgoose, G. R., and Kalma, J. D.:
Three-dimensional soil moisture profile retrieval by assimilation
ofnear-surface measurements: Simplified Kalman filter
covarianceforecasting and field application, Water Resour. Res.,
38, 1–18,2002.
Warrach-Sagi, K., Wulfmeyer, V., Grasselt, R., Ament, F., and
Sim-mer, C.: Streamflow simulations reveal the impact of the
soilparameterization, Meteorol. Z.,17, 751–762, 2008.
Wulfmeyer, V., Behrendt, A., Bauer, H.-S., Kottmeier,
C.,Corsmeier, U., Blyth, A., Craig, G., Schumann, U., Hagen,
M.,Crewell, S., Di Girolamo, P., Flamant, C., Miller, M.,
Montani,A., Mobbs, S., Richard, E., Rotach, M. W., Arpagaus, M.,
Russ-chenberg, H., Schlüssel, P., K̈onig, M., G̈artner, V.,
Steinacker,R., Dorninger, M., Turner, D. D., Weckwerth, T., Hense,
A., andSimmer, C.: The Convective and Orographically-induced
Precip-itation Study: A Research and Development Project of the
WorldWeather Research Program for improving quantitative
precipita-tion forecasting in low-mountain regions, B. Am.
Meteorol. Soc.,89/10, 1477-1486 , doi:10.1175/2008BAMS2367.1,
2008.
Geosci. Model Dev., 3, 1–12, 2010
www.geosci-model-dev.net/3/1/2010/
http://www.hydrol-earth-syst-sci.net/9/586/2005/